PSI - Issue 84

Marianna Crognale et al. / Procedia Structural Integrity 84 (2026) 898–905

903

Fig. 4. Fiber-computed versus ML-predicted moment degradation factor as a function of curvature for a representative degraded pier section (C2–S0, =15 MN). Table 2. Performance of the machine-learning surrogate for section degradation prediction. Quantity R² MAE η E (stiffness factor) 0.999 0.0013 η M (moment factor) 0.999 0.0017 It must be noted that the validation scenario was applied to the training data only, then, in a limited and controlled setting, primarily for preliminary debugging and fidelity checks. In this context, the objective was to verify the correctness of the implementation and the model capacity to fit the given data, rather than to assess generalization performance. Accordingly, the reported results should not be interpreted as an estimate of predictive validity nor used for model selection. 3.1. ML-Informed Integration within Pushover Analysis The trained ML surrogates enable a deterioration-aware interpretation of global pushover results without resorting to fully fiber-based system-level modeling. During the pushover analysis, curvature demand at critical pier sections is monitored and converted into an equivalent ductility measure. Based on the current curvature demand and axial force level, the ML surrogate provides corresponding degradation factors and . In other words, rather than directly modifying fiber discretization during the nonlinear solution process, the ML predictions are used to interpret and post-process the evolution of stiffness and strength degradation at the section level. This strategy avoids numerical instability associated with online section rebuilding while still providing physically informed insight into progressive material deterioration during increasing load. The resulting pushover capacity curves therefore reflect not only geometric and global nonlinearities, but also the degradation trends inferred from section-level physics through the ML surrogate. This hybrid approach retains computational efficiency while enhancing the interpretability of global seismic performance under material degradation. Fig. 5 shows the global pushover response of the viaduct (base shear versus control displacement) for a representative degraded material state (C2–S0), color-coded by the ML-inferred section stiffness degradation factor at the identified critical pier section. The pushover curve itself is unmodified. The color scale provides a post processing interpretation of progressive sectional stiffness degradation along the pushover path. The consistency of the results appears to be excellent.